Enhanced YOLOv8 for Industrial Polymer Films: A Semi-Supervised Framework for Micron-Scale Defect Detection

Xiaoxia Yu¹Bingyu Hu^2,3Weifeng Jiang⁴Jinru Wan¹Xinduoji Yang²Nianbo Liu^2,3Xiaoyan Dong^1*

• ¹Zhejiang Juhua Co Ltd, Quzhou, China
• ²University of Electronic Science and Technology of China School of Electronic Science and Engineering, Chengdu, China
• ³University of Electronic Science and Technology of China Yangtze Delta Region Institute Quzhou, Quzhou, China
• ⁴Juhua Group Corporation, Quzhou, China

Polymer material films are produced through extrusion machines, and their surfaces can develop micro-defects due to process and operational influences. The quantity and size of these defects significantly impact product quality. As traditional machine learning defect detection methods suffer from low accuracy and poor adaptability to complex scenarios, requiring extensive effort for parameter tuning and exhibiting weak generalization capability, this paper proposes an improved YOLOv8 method to identify micro-defects on films. The approach embeds the CBAM attention mechanism into high-level networks to address feature sparsity in small target detection samples. Simultaneously, given the difficulty in obtaining large annotated datasets, we employ the Mean Teacher method for semi-supervised learning using limited labeled data. During training, the method optimizes neural network gradients through an improved loss function based on normalized Wasserstein distance (NWD), mitigating gradient instability caused by scale variations and enhancing detection accuracy for small targets. Additionally, a proposed multi-threshold mask segmentation algorithm extracts defect contours for further feature analysis. Experimental results demonstrate that the improved YOLOv8 algorithm achieves an 8.26% increase in mAP@0.5 compared to the baseline. It exhibits higher precision for small targets, and maintains defect detection rates exceeding 95.0% across validation data of varying image sizes, thereby meeting industrial production requirements. In generalization validation, the model demonstrates superior performance compared to traditional methods under test environments with lighting variations and environmental contamination.